Inspiration
Digital scams are becoming more convincing as generative AI makes it easier to produce realistic messages, images, job offers and synthetic voices. However, most consumer fraud-awareness tools still rely on posters, warning lists or videos. These approaches explain scams, but they do not test whether someone can make the right decision while facing urgency, authority or emotional pressure.
ScamShield Arena was inspired by a simple idea: scammers use generative AI to rehearse attacks against society, so we should use generative AI to help society rehearse its defence.
What it does
ScamShield Arena is a free, multilingual scam-resistance training platform for students, parents, senior citizens, job seekers, online shoppers and small businesses.
Users choose a profile and enter a clearly labelled, fictional simulation. The current MVP includes:
- A UPI payment trap
- A family-emergency synthetic voice scam
- A fake job recruitment scam
During each challenge, the user must decide how they would respond. ScamShield then evaluates the decision, identifies the manipulation techniques used, explains the associated risk and provides a short corrective lesson.
Users receive a private protection score that measures their progress across payment fraud, identity theft, emotional manipulation and independent verification.
How we built it
The interactive frontend was built using React, TypeScript and a Next.js-compatible Vite environment. It provides profile selection, multilingual controls, three complete simulations, decision evaluation, progress tracking and responsive accessibility.
The backend was built with FastAPI and Pydantic. It supports a credential-free demonstration mode and a live generative-media mode.
In live mode, Genblaze orchestrates the multimodal workflow. It can generate a fictional visual simulation and a generic synthetic voice lesson through GMI Cloud providers. Genblaze manages the pipeline, retries and output provenance.
Backblaze B2 acts as the durable system of record. Generated images, audio, safety metadata, anonymous learning results and canonical provenance manifests are organized under a shared training-run prefix. This allows every media asset to remain connected to the model, prompt policy, parent scenario and integrity record that produced it.
How we use Genblaze
Genblaze is the orchestration layer connecting adaptive training decisions to generative-media providers.
The pipeline:
- Receives the learner profile and weakest verification skill.
- Applies a strict safety policy.
- Generates fictional image and synthetic audio training assets.
- Retries unsuccessful generation steps.
- Sends completed assets to Backblaze B2.
- Produces a canonical SHA-256 provenance manifest.
The provider can be changed without rewriting the behavioural-training workflow.
How we use Backblaze B2
Backblaze B2 is more than an output-image host. It stores and organizes:
- Generated training images
- Synthetic audio lessons
- Safety-policy metadata
- Anonymous training outcomes
- Scenario lineage
- Canonical Genblaze manifests
- Content integrity records
The object structure groups all outputs under scamshield/training-runs/<run-id>/. B2 lifecycle policies can remove temporary intermediate assets while preserving reviewed simulations and their provenance records.
Safety by design
Because the project generates scam-awareness media, safety boundaries are built into both the interface and generation pipeline.
ScamShield:
- Uses fictional identities only
- Never impersonates real people
- Prohibits voice cloning and uploaded voice samples
- Uses no real banks or company branding
- Produces no functional payment links or QR codes
- Does not collect credentials or payment information
- Permanently labels content as a training simulation
- Does not allow raw deceptive media to be exported
- Stores generation and safety provenance with every asset
Challenges we faced
The main challenge was creating realistic learning experiences without producing reusable scam materials. We addressed this by separating behavioural learning objectives from media generation and applying strict constraints before every generation step.
Another challenge was ensuring that the application remained testable without paid provider credentials. We created a deterministic demonstration mode while keeping the production Genblaze and B2 pipeline available for live generation.
We also had to balance realism with accessibility. The interface avoids fear-heavy cybersecurity imagery and instead uses clear language, large controls, immediate explanations and visible safety labels.
Accomplishments that we are proud of
- Built three complete interactive scam simulations
- Designed the platform for six different social groups
- Measured user decisions instead of presenting passive awareness content
- Added adaptive scoring and personalised corrective lessons
- Created a real Genblaze image-and-audio pipeline
- Designed meaningful Backblaze B2 media and provenance storage
- Implemented explicit safeguards against harmful media generation
- Created a responsive, multilingual and credential-free demonstration
What we learned
We learned that effective fraud education should focus on behaviour rather than memorisation. Knowing that scams use urgency is different from successfully resisting urgency during an interaction.
We also learned that media provenance is essential for responsible generative applications. Generated assets should not be separated from the policies, prompts, models and lineage that created them.
What's next for ScamShield Arena
Future development will include:
- Additional regional languages
- More scam categories and difficulty levels
- Spaced-retention assessments
- Community facilitator and NGO dashboards
- Reviewed scenario packs from banks and digital-safety organisations
- Family safe-word training
- Anonymous population-level impact analytics
- Tamper-evident B2 Object Lock manifests
- WhatsApp-friendly corrective lessons
Our goal is to make practical scam-resistance training freely accessible before fraud reaches the user.
Built With
- accessibility
- ai
- b2
- backblaze
- cloud
- cybersecurity
- fastapi
- fraud
- genblaze
- generative
- gmi
- html5
- multimodal
- next.js
- prevention
- pydantic
- python
- react
- s3-compatible
- safety
- storage
- typescript
- vite
Log in or sign up for Devpost to join the conversation.